News from the IT community on October 13th: Developer Morris Hoffman ( Maurice Heumann ) released a project report on October 9th, stating that the team spent 3 months and over 500 billion yuan ( Tokens ) to decompile the AI agent for the 2009 game "Call of Duty: Modern Warfare 2" and rebuild it into C++ source code.

In terms of team members, this project was led by Hoffman and received assistance from RektInator, Future, st0rm, and other community members.

In terms of calling AI, the team initially used both Claude Max (20x) and Codex Pro for subscription, while running the AI agents in Claude Code CLI and Codex CLI. The models covered Sonnet 5, Opus 5.5, Luna, Sol, and Terra. The agents tracked tasks through GitHub Issues and exchanged information via Discord.

In the first month of the project, 4 agents were deployed. Three of them were responsible for decompiling and submitting code, while one was in charge of reviewing it. During this phase, approximately 80% of the functions were completed, allowing the game to start, display the main menu, and load maps. However, the team discovered that although the code was readable, there were semantic errors at the levels of function signatures, types, structural layout, and logic.

Subsequently, the team introduced a "byte-by-byte matching decompilation" verification mechanism. This mechanism rebuilds the target files using the original game compiler and compares the rebuilt functions byte by byte with those in the original EXE / PDB. Those that pass the comparison are marked as PASS, while those that do not are modified again by the agent.

To prevent agents from bypassing verification, the team prohibits inline assembly, naked functions, object patches, and the embedding of bytes in code. Instead, they use CI to compare the hash values of verification scripts with GitHub Actions and secret. This mechanism enables low-cost models to complete high-precision tasks under strict feedback, thereby replacing manual or agent review processes.

In the later stages of the project, 14 Luna agents and 2 Opus 5.5 agents were primarily in use. These agents submitted code through independent branches and Pull Request. Ultimately, 99% of the game functions were incorporated into the reconstructed source code, with 83% of the functions achieving byte-for-byte matching; however, the remaining functions could not be fully matched due to non-deterministic characteristics or linker behavior.

Byte-by-byte matching ( Byte-for-byte Matching ) refers to comparing the binary data of the reconstructed function after compilation with the corresponding function in the original program byte by byte. If the data is consistent, it usually indicates that the function retains the same semantics and behavior as in the original version in the target compilation environment.

Hoyman stated that the relevant code will not be made public, and the results of the reconstruction will only be retained for private use.

IT's home comes with reference links: 500+ Billion Tokens Later : Letting AI Agents Decompile A First - Person Shooter.